Weight sensitivity analysis of neural network based manipulator controllers

Manipulator control has been an active research topic for some time. Without a parametric model of robot dynamics, learning control techniques have been viable alternatives for repeated trajectory following tasks. Specifically, neural network based algorithms have been the focus of such research. Although these techniques were shown to work effectively in simulation experiments, coupled and nonlinear nature of parameter update dynamics makes an effective mathematical analysis difficult. This paper studies the weight sensitivity and convergence properties of artificial neural network based robotic controllers. The results obtained help us in understanding the localized stability properties of the closed loop nonlinear system dynamics.